Evidence Accumulation in Multiobjective Data Clustering

Multiobjective approaches to data clustering return sets of solutions that correspond to trade-offs between different clustering objectives. Here, an established ensemble technique (evidence-accumulation) is applied to the identification of shared features within the set of clustering solutions returned by the multiobjective clustering method MOCK. We show that this approach can be employed to achieve a four-fold reduction in the number of candidate solutions, whilst maintaining the accuracy of MOCK’s best clustering solutions. We also find that the resulting knowledge provides a novel design basis for the visual exploration and comparison of different clustering solutions. There are clear parallels with recent work on ‘innovization’, where it was suggested that the design-space analysis of the solution sets returned by multiobjective optimization may provide deep insight into the core design principles of good solutions.